Distributed Bandit-Based Cooperative Coevolution for Large-Scale Multi-Objective Data Publishing

Yong-Feng Ge, Hua Wang, Elisa Bertino, Jinli Cao, Yanchun Zhang, Zhonglong Zheng

Abstract: When publishing data, organizations must balance transparency requirements with data privacy concerns. This paper introduces a Distributed Bandit-based Cooperative Coevolution (DBCC) algorithm that addresses scalability and adaptability challenges in large-scale datasets. The DBCC optimizes privacy protection and transparency through a divide-and-conquer approach using three sequential modules: clustering-based grouping module, bandit-based Cooperative Coevolution (CC) module, and evolving module. In the grouping module, the DBCC algorithm clusters similar records into the same subproblems. The CC module optimizes subproblems in sub-populations simultaneously while measuring their current and potential contributions. It adaptively allocates population resources to maintain a balance between exploration and exploitation. To enable Differential Evolution (DE) in discrete-domain data publishing, the algorithm implements a set-based mutation operation and a transparency-driven crossover operation. The evolving module then generates the final complete solutions. Through our experiment, we demonstrate that the proposed DBCC achieves significantly higher solution accuracy than existing state-of-the-art algorithms across all 15 test instances, showing an improvement of 247.16%. We validate the effectiveness of all DBCC components and investigate the impact of parameters. Additionally, our investigation reveals that speedup ratios can reach up to 3.45 when parallel granularity is 10.

External IDs:dblp:journals/tsc/GeWBCZZ25